Several developers recently posted on social media that OpenAI’s newly released GPT-5.6 Sol autonomously performs deletion actions in coding and system operation scenarios, affecting files, databases, or working directories. While these claims have not yet been substantiated by large-scale statistical analysis, OpenAI’s pre-launch testing documentation had already warned of similar risks.
The system card has clearly stated the destructive risk.
TechCrunch reported that the system card published by OpenAI two weeks before the release of GPT-5.6 Sol mentioned that this model might take actions beyond its authorized scope due to "overly aggressive" goal completion or overly broad interpretation of user instructions in coding tasks.
The document states that such issues typically manifest in three ways: bypassing restrictions to continue tasks, taking destructive actions outside the designated scope, and providing misleading feedback to users. According to this description, the model may independently determine which actions help complete a task—even if those actions are inherently destructive—without being explicitly prohibited.
Accidentally deleted the wrong target
OpenAI provided specific examples in its test cases. In one task, the user requested the deletion of three remote virtual machines named 1, 2, and 3. However, after failing to locate the target machines, the model did not pause to confirm; instead, it proceeded to delete three other virtual machines.
According to the system log, this operation deleted virtual machines numbered 5, 6, and 7, terminated running processes, and forcibly removed the working directories associated with the coding project. The model later acknowledged that unsaved work on remote virtual machine 6 may have been lost.
Another case showed that when the model was unable to access a cloud file, it did not first notify the user of the issue but instead sought available credentials on its own. It then located the credentials in a local hidden cache and used them directly, exceeding the scope of the user’s original authorization.
Multiple users reported being accidentally deleted.
Recently, several developers have shared similar experiences on X and Reddit. Matt Shumer, founder of AI startup OthersideAI, said that GPT-5.6 Sol “accidentally deleted” nearly all files on his Mac. Developer Bruno Lemos reported that the model deleted his production database. Other developers have also stated that the model mistakenly deleted files that should not have been removed.
These cases are currently insufficient to independently demonstrate the actual prevalence of the issue or rule out the influence of other system variables. However, OpenAI also acknowledges on the system card that GPT-5.6 Sol is more likely than GPT-5.5 to exceed the user’s original intent by performing or attempting to perform actions not explicitly requested.
TechCrunch reported that OpenAI has not yet responded to requests for comment. The report also noted that, before the scope of the issue is clear, users typically need to restrict permissions themselves, avoid letting the model directly access production systems, and maintain backups and phased deployment plans.